step_04000 — SpatialAV2AV checkpoint

Latest & best-tested full fine-tune weights for LTX-2.3 (22B) spatial audio-video editing.

Provenance

Field	Value
Source path	`/apdcephfs_zwfy11/share_305172035/helensliang/Projects/LTX-2-SpatialAV2AV/outputs_SpatialAV2AV_train/2026.06.19-17.33.11/checkpoints/model_weights_step_04000.safetensors`
Training run	`2026.06.19-17.33.11`
Step	4000 / 20000 (checkpoints saved every 1000 steps)
Checkpoint written	2026-06-22 14:05
File size	37,979,154,726 bytes (~36 GiB)
Base model	`LTX-2.3/ltx-2.3-22b-dev.safetensors` (full-parameter fine-tune, FSDP FULL_SHARD, 8×H20)
Training time to this step	~341 h wall (60 s/step)

Why this checkpoint

Last checkpoint of the run and the one used in the final inference tests (test8_step4000_* / realtest8_step4000_*, run 2026-07-05). Video-dominant loss stayed low and stable; stereo L≠R verified on decode.

Loss (video weight 0.85 / audio weight 0.15)

Step	Loss	v_loss
1000	0.1089	0.0937
2000	0.1528	0.1555
3000	0.1099	0.1051
4000	0.1191	0.1221

Key training config

learning_rate: 1.0e-05, steps: 20000, batch_size: 1, gradient checkpointing on
timestep_sampling_mode: shifted_logit_normal
loss_video_weight: 0.85, loss_audio_weight: 0.15, with_audio: true, use_ref_audio: false
edit_sample_n_frames: 113, edit_min_size: 288
resolution_buckets: 288x384x113; 384x288x113; 320x320x113; 352x352x113
training_strategy.name: spatialav2av
data: SpatialAV2AV_full/all.list (source→edit stereo pairs)

⚠️ Format quirk — read before loading

This .safetensors file is actually a torch.save (zip/pickle) file, not safetensors. Magic bytes = PK\x03\x04.

❌ safetensors loaders fail with HeaderTooLarge.
✅ torch.load() + transformer.load_state_dict(..., strict=False) → missing=0 unexpected=0.

Use SpatialAV2AV_inference.py --full_weights <this file> (it handles the injection).